Sales in the age of AI

Is human review the bottleneck in AI-assisted sales? What Anthropic's own workforce data suggests

By Uassist.ME Editorial Team · Published by Uassist.ME · Published 2026-10-20 · Last updated 2026-10-20

Disclosure: ThePipelineBots is published by Uassist.ME (uassistme.com), which sells outsourced sales-support and CRM teams. Some of that work is reviewing and acting on what software produces, so we have a stake in the argument that human review matters. Weigh our take with that in mind. Every figure below links to its source.

Short answer: At Anthropic, "a company at the frontier of AI adoption," its own data says human attention is the bottleneck: interactive Claude sessions spend about 80% of their time idly waiting for human input. That's company-wide data from one AI vendor, not a study of sales teams. Anthropic also reports that one account executive's email and CRM automations replaced around 20% of their prior tasks, and that another used similar automations to talk to 30% more of their customers. Vendors are now designing around the human step too, with live handoffs to reps and separate systems that supervise agents.

What did Anthropic find about its own salespeople? (The good)

On Oct 16, 2026, Anthropic published a report on how AI has changed its own workforce. It draws on internal data on employment, AI use, output, surveys, calendars and hiring records, plus 17 employee interviews (Anthropic, Oct 16, 2026). SaaStr summarized it on Oct 19 (SaaStr, Oct 19, 2026). The figures below come from Anthropic's report, not the summary.

Caveats (ours, plus the report's own): our caveat is that the AE figures are interview anecdotes, not measured sales metrics. The report itself says "part of this pattern could be driven by representatives with more sales opportunities selecting into automation." It also notes that "it is easier to close deals when you sell a strong model," during a period of "explosive business growth." Anthropic sells AI, so this is a vendor reporting on its own use of its own product.

Where does the human bottleneck show up? (The ugly)

Anthropic's report is direct about it: "human attention is a constant bottleneck" (same source).

That matches what G2 found in its July 2026 survey: 31% of B2B organizations spend 6 or more hours a week overseeing agents. We covered that in Before you let an AI agent touch your CRM: what should you check?

Are vendors building the human step back in? (The good, vendor-reported)

monday.com's Harmony launched on Oct 19, 2026 as a voice AI agent platform for sales teams (monday.com IR, Oct 19, 2026). monday.com says its agents run calls from prospecting to "warm transfers to live reps with a full briefing." On Harmony's FAQ, the company says "a human closes," and "by default the agent tells people it is an AI," a setting customers can change (Harmony.ai, accessed Oct 20, 2026).

In the release, Richard Stubbs of UK car-finance broker Bright Car Finance says the agent "qualifies about half of the conversations it has and hands 80% of those live to our advisers, driving a 9% increase in deals closed in less than a year." That's a customer quote in a vendor press release, with no baseline published. monday.com also claims "a contact rate of up to 70%" without saying how it was measured. No pricing was published.

Should the agent's supervisor be separate from the agent? (The ugly)

Microsoft thinks so. In a post on Oct 19, 2026, Microsoft Security EVP Hayete Gallot wrote: "Unfortunately, frontier models, regardless of their origin, cannot be trusted." She said agents "are frequently overprivileged" and called for "Guardian Agents" that "watch and moderate every action of every agent, before it happens," as part of Microsoft's Project Perception (Microsoft, Oct 19, 2026). She wrote: "Agentic security provides speed. Practitioners provide judgment."

According to TechCrunch's reporting, Microsoft also added agent detection to Microsoft Defender for "up to 60 types of agents" that employees bring in. It also added task-based permission limits, including "requiring human approval if an agent appears to be making risky decisions" (TechCrunch, Oct 19, 2026). TechCrunch also quotes Satya Nadella's own blog: "we need to separate the supply of intelligence from the authority over it." TechCrunch calls Microsoft's move "self-serving," since Microsoft sells the monitoring layer. Microsoft's post isn't specific to sales or CRM agents; that link is our angle.

How can a sales team plan for review capacity?

All of these are our suggestions, drawn from the sources above. They aren't vendor guidance.

  1. Count review time as part of the cost. Before adding an agent, estimate who will read and approve its output, and for how many hours a week. (Our suggestion.)
  2. Name one human owner per agent, who approves its changes and can pause it. (Our suggestion.)
  3. Review by exception. Have people check flagged or high-risk actions, plus a sample of the rest, instead of reading everything. (Our suggestion.)
  4. Put a human at the handoff that matters, such as a live transfer to a rep for qualified leads rather than letting the agent close. (Our suggestion.)
  5. Keep approvals for destructive or bulk CRM changes, and keep the agent's permissions narrow. (Our suggestion.)
  6. Write decisions down where agents can read them. Anthropic says that when a meeting ends with no notes or report-out, "agents are missing crucial context." (Our suggestion, based on the report.)
  7. Track how long agents wait on people. If approvals pile up, add reviewers or cut the agent's volume. (Our suggestion.)

Related: Do you need to clean your CRM before using an AI sales agent?

FAQ

How much time do AI agents spend waiting for humans? In Anthropic's internal data, interactive Claude sessions spend around 80% of their time idly waiting for human input (Anthropic, Oct 16, 2026).

Did AI help Anthropic's salespeople reach more customers? Anthropic reports that one account executive's automations replaced around 20% of their prior tasks, and that another used similar automations to talk to 30% more of their customers. These are Anthropic's paraphrases; our caveat is that they're interview anecdotes, not measured metrics.

What is "meta-work"? Anthropic's term for managing agents: writing specs, having one agent review another's output and monitoring pipelines. It grew from 10% to 22% of the average employee's interactive Claude sessions between fall 2025 and spring 2026.

Does monday.com's Harmony hand calls to human reps? monday.com says Harmony warm-transfers qualified prospects to live reps with a full briefing, and that by default the agent tells people it's an AI. These are vendor claims.

Why does Microsoft say AI agents should be supervised separately? Microsoft Security's Hayete Gallot wrote that frontier models "cannot be trusted" because they're trained to attain results, which can lead to reward hacking. She added that agents are unpredictable and frequently overprivileged, and proposed Guardian Agents that monitor every agent action before it happens.

Sources (opened 2026-10-20)

  1. Anthropic (Nathan Wilmers, Regina Catipon, Maxim Massenkoff, Peter McCrory), "How employment is changing at Anthropic," Oct 16, 2026. Vendor reporting on itself. https://www.anthropic.com/research/employment-at-anthropic
  2. SaaStr (Jason Lemkin), "5 Interesting Learnings From Anthropic's Data on Its Own Workforce…," Oct 19, 2026. Secondary summary; used for context only. https://www.saastr.com/5-interesting-learnings-from-anthropics-data-on-its-own-workforce-junior-technical-hiring-down-from-27-to-9-agents-working-6x-the-hours-of-humans-and-agents-waiting-on-a-human-80-of-the-time/
  3. monday.com investor relations, "Voice AI Platform Harmony Launches to Convert Customer Conversations into Revenue," Oct 19, 2026. Vendor-reported. https://ir.monday.com/news-and-events/news-releases/news-details/2026/Voice-AI-Platform-Harmony-Launches-to-Convert-Customer-Conversations-into-Revenue/default.aspx
  4. Harmony.ai homepage and FAQ, undated, accessed Oct 20, 2026. Vendor site. https://harmony.ai/
  5. Microsoft (Hayete Gallot), "Securing the agentic insider," Oct 19, 2026. Vendor (sells agent security). https://blogs.microsoft.com/blog/2026/10/19/securing-the-agentic-insider/
  6. TechCrunch (Julie Bort), "Microsoft says AI agents are like insider threats as 'frontier models can't be trusted'," Oct 19, 2026. https://techcrunch.com/2026/10/19/microsoft-says-ai-agents-are-like-insider-threats-as-frontier-models-cant-be-trusted/
  7. G2, "Agentic Chasm: 2026 AI Agents Report," last updated Oct 14, 2026 (via our Oct 19 article). Interested party. https://sell.g2.com/g2-2026-ai-agents-insight-report